Insurance companies (and many other types of businesses) send out large numbers of quotations to prospective customers, and success hinges on encouraging a large proportion of these customers to accept these offers as soon as possible. This article describes an approach we have been thinking about for doing just that.

Framing the Problem

The problem starts with quotes, which offer a service (such as an insurance policy) for a fixed price. The quote has an expiry date, typically 30-90 days from the date the quote is generated (inception).

The company may wait passively for customers to accept the quotes, or for the quotes to expire. Or, it may try interventions to encourage customers to accept quotes. For example, by sending e-mail or SMS reminders, changing the price, bunding other items into the offer, and so on.

In any case, the amount of time it takes for customers to accept the quote (inception period) is likely to be distributed something like this (plot of randomly generated synthetic data set):

There are two ways to add value here:

  1. Move the curve up, by getting more people to accept
  2. Shift the curve to the left, by getting people to accept earlier

This must be done with an eye to costs -- the value added must be greater than the cost of capturing it.

Marketing Actions: Mining Response Curves

As noted above, there are any number of ways you could encourage customers to accept a quote, including:

As soon as we try any of these interventions, we start accumulating data on the success of these interventions, both in terms of success rate (percent of targets who convert), and the time it takes for them to respond. Mining this data, we can build up response curves such as the following (screen shot of ficticious marketing action parameters, from our web prototype):

In the first of the three hypothetical examples above, the response curve for E-mail reminders shows that an follow-up e-mail has a peak effect of 0.35% conversion, about six days after the quote is offered. It is lower before that, and declines slowly after the peak, reaching zero at 12 days. All this at a cost of 0.1 pence (0.001 GBP) per e-mail.

Over time, with the right data, we can measure the response behaviour, and build up these curves, for any interventions tried, and for any sub-segment of customers targetted.

Using these Response Curves to Optimize Marketing Spend

The response curves are interesting in their own right, but are of limited use without a tool to apply them to marketing planning. The problem is that, on their own, they do not tell you where to apply marketing spend. The peak of each response curve is only part of the story, since it is the age of each quote that helps determine the best allocation of spend, across the basket of quotes that are awaiting customer action.

We are therefore proposing an iterative process of optimizing marketing spend against the outstanding quotes at any time. This process would work as follows:

  1. Periodically (e.g., every hour or every day), we look at all outstanding quotes
  2. Given the age of each quote, and the response curves we have learned about different marketing acions, we apply the action with the highest expected value, i.e., with the highest probability of success on that day, subject to the action's cost
  3. The choice of action could also take into account the previous history for each quote, e.g., e-mail reminders are not sent to the same customer twice is short succession
  4. We use an optimization algorithm to make these decisions subject to a daily or weekly cap on marketing spend, if there is one
  5. We do this continually, reflecting the changing pool of quotes outstanding (because some will have been accepted, and new ones will have been issued), and the changing response curves

What this process achieves is taking into account the timing of marketing actions with respect to the lifecycle of each quote, and the changing characteristics of each possible marketing action over time.

Impact of this Approach

We have built a working prototype of this approach, and applied it to a synthesised data set. Simulating the approach on this data set showed the following:

These results are obviously subject to the data, and to the parameters of the response curves, but nevertheless suggest that there is signicant possible value in applying this approach to real marketing activities in an insurance company.

Add a comment